[tannevaled]

Closes #27 — all nine parse errors and the mis-parse, plus the two files the
issue deliberately left unclaimed, plus four further gaps each of them uncovered.

Every one of the 68 files in the pinned ruby/spec language/ corpus now
parses (13 did not before), and so do 4405 of the 4413 .rb files in the
whole corpus, up from 4391 — the 8 that remain include two deliberate
syntax-error fixtures. No file that parsed before stopped parsing.

Each fix reads the MRI rule first; the commit messages cite it by name and line
in parse.y v3_4_0, and every behaviour was checked against the installed MRI
4.0.5, comparing values through rbgo rather than merely "it parses".

What the issue filed

#	source	rule read in parse.y v3_4_0	commit
1, 2	alias :"#{:a}" v / undef :"#{:m}"	`fitem: fname	symbol, symbol: ssym
3	(o).a, (o).a = 1, 2	mlhs_node: primary_value call_op tIDENTIFIER (3639-3652) — the receiver is any primary_value	ac7e3a1
4	defined?($zz and true)	keyword_defined '\n'? '(' begin_defined expr rparen (4458), a primary; and keyword_defined '\n'? begin_defined arg (4039)	c1df337
5	for i, in [[1,2]]	`for_var: lhs	mlhs, with mlhs_head: mlhs_item ','` a complete mlhs
6	p({:a==>1})	parse_ident: `c=='=' && … (!peek('=')
7	in Array(0,1,2)	p_expr_basic: p_const p_lparen p_args rparen — the same quartet as p_lbracket	afab703
8	p $=	parse_gvar's tGVAR arm lists _ ~ * $ ? ! @ / \ ; , . = : < > "	05f6a4c
9	-> a=a() { a }	lambda_beginning_p() = lpar_beg == paren_nest, set at the ->	9b7c51b
mis-parse	o.s (:a){ 1 }	tLPAREN_ARG … {SET_LEX_STATE(EXPR_ENDARG);} ')' (4405) then command: … command_args cmd_brace_block (3464)	1ba2c43

The mis-parse is the one that mattered most: it was silent. o.s (:a){ 1 }
bound the block to :a, so rbgo raised LocalJumpError where MRI prints 1. The
negative case is pinned too — f g(1) { 2 } uses a HUGGING paren, so its )
leaves EXPR_END and MRI gives the block to g(1); that still happens.

The two files the issue could not reduce

Both reduce, and neither error was at the line it pointed to in the way it
looked:

• class_spec.rb (and metaclass_spec.rb, same cause): class C; end.should == nil.
  The parser let a definition take a .method and a modifier but not be an
  OPERAND. In MRI k_class cpath superclass bodystmt k_end, the class << expr
  form, k_module cpath bodystmt k_end and defn_head f_arglist bodystmt k_end
  are all alternatives of primary (4599-4657). Removing the statement-level
  interception was the whole fix (9ce5aec).
• string_spec.rb: %@hey #{@ip}@. The %-literal terminator scan looked
  INSIDE the interpolation and closed on the @ of @ip. MRI's tokadd_string
  leaves the scan at #{ and hands the embedded expression to the main lexer.
  Two more %-delimiter faults were behind it: %=hey= (MRI tests IS_BEG()
  before the op-assign =, so = delimits a literal where a value is expected
  and stays op-assign after one) and %\hey\ (the terminator test precedes the
  backslash test in tokadd_string) (92c7bd2).

Found on the way, each blocking the next line of a spec file

• defined?(break), defined?(return): a bare break/next/retry is a
  primary (4682-4697) and so is a bare return (k_return, 4438), with the
  call_args forms alternatives of command (c1df337).
• in [0, 1, ] parsed but silently became the EXACT pattern [0, 1], so a
  case [0,1,2,3] fell to else where MRI matches. p_top_expr_body: p_expr ','
  makes a trailing comma an ANONYMOUS REST (afab703).
• in {"a": 0}: p_kw_label: tLABEL | tSTRING_BEG string_contents tLABEL_END.
  in {"a" => 1} stays refused, as ruby/spec requires (afab703).
• in ^@a: p_var_ref: '^' tIDENTIFIER | '^' nonlocal_var (afab703).
• |a, (b, (c, d))|: f_marg: f_norm_arg | tLPAREN f_margs rparen nests, but
  parseBlockParams held a flat copy of logic the def list already had
  recursively (1ba2c43).

AST changes, all additive

ast.Alias gains NewNameExpr/OldNameExpr, ast.Undef gains Exprs,
ast.Block gains Locals, ast.For gains Target. Each is nil/empty for the
shapes that worked before, so existing field reads keep their meaning and
existing consumers keep building. ast.For.Target is the one a consumer must
learn: Vars is EMPTY whenever it is set, so there is only ever one place to
read the loop variable from, and for i in … / for a, b in … still arrive in
Vars unchanged.

Measurement

rbgo built against this branch, over the pinned corpus, before and after.
Ten files went from "no result at all" to a measured result, revealing
311 passing examples that no measurement had ever seen:

file	now
pattern_matching_spec.rb	pass=95 fail=16 error=2
predefined_spec.rb	pass=68 fail=89 error=4
class_spec.rb	pass=32 fail=8 error=3
hash_spec.rb	pass=29 fail=11 error=0
string_spec.rb	pass=23 fail=9 error=2
alias_spec.rb	pass=22 fail=2 error=2
metaclass_spec.rb	pass=18 fail=0 error=3
assignments_spec.rb	pass=18 fail=20 error=0
undef_spec.rb	pass=6 fail=0 error=2
lambda_spec.rb	pass=0 fail=0 error=2 (its SpecEvaluate helper is missing from the corpus harness)

Corpus totals: pass 1031 → 1342, files with no result 14 → 4. Nothing regressed:
the before/after diff over all 68 files contains only these additions.

The four files still reporting nothing all PARSE now and stop in rbgo's
compiler, not here:

• for_spec.rb — undefined local variable "j": rbgo reads For.Vars, which is
  empty for for i, *j in …. It needs to read For.Target.
• block_spec.rb — undefined local variable "bl": block-locals need a slot.
  Block.Locals now carries the names (74ed65f); rbgo has to allocate them.
• defined_spec.rb — undefined local variable "a": defined?(a += 1) declares
  the local at parse time, as MRI does, but rbgo's compiler allocates no slot for
  it. Reproduces identically on main's parser, so it is pre-existing.
• variables_spec.rb — cannot compile *ast.SplatArg.

One more consumer-side finding, from assignments_spec.rb: rbgo evaluates the
right-hand side of a multiple assignment before the target receivers, where MRI
evaluates the receivers first. The AST records the two sides, not an order, so
that is a compiler matter.

Gates

go build ./..., go vet ./..., go test -race ./... all green; the 100%
coverage gate holds (and peek3's end-of-input branch, which the ROUNDED total
was hiding at 100.0%, is now covered — 60303c5). gofmt -l . lists only
paren_singleton_def_test.go, which is untouched and already listed on main.

🤖 Generated with Claude Code